Research On Hierarchical Optimization Semantic Segmentation Algorithm Based On Adjacency Dependence

Semantic segmentation has a wide range of applications in autonomous driving,medical analysis,and computational photography due to a comprehensive scene description of object category,location,and shape information.This paper focuses on the semantic segmentation algorithm based on deep learning.Although the introduction of full convolutional networks makes deep learning methods far more accurate than traditional algorithms,there are also some problems inherent in the network that have not been overcome so far,especially the contradictions between high-level semantics and spatial information.This paper attempts to bridge the gap between low-level and high-level features in order to optimize the problem that the upper segmentation is too rough,thereby significantly improving the segmentation quality.The research content of this article can be summarized as follows:Firstly,aiming at the problem of small resolution of the feature map caused by the continuous downsampling of the network to extract abstract features,this paper uses the atrous convolution with different expansion rates to replace the ordinary convolution operation in the back end of the original network,which maintains the size of each neuron's receptive field with the ability of the network to extract abstract features and the resolution of the segmentation result is not too low.Based on this,in view of the disadvantages of independent prediction of image pixels by general methods,this paper proposes the concept of Hierarchical Adjacency Dependence by observing the fact that pixels with similar local features tend to be semantically consistent.Considering the dependencies between local pixels in the feature layers of different abstract expressions,this paper change the independent prediction of pixels to the joint prediction of surrounding pixels to make the prediction result more compact.This achieves the purpose of restoring segmentation detailsSecondly,aiming at the combination optimization problem of different levels of features,this paper studies the difference in the effect of image optimization on the features of each layer.The branches are strategically combined to further optimize the segmentation results by cascading the layers that are manually selected,so that the effect of repairing is superimposed.Finally,in order to automate feature branch selection,this paper introduces a pixel-level confidence strategy to allow the network to autonomously select the appropriate feature layer during the training process.The problem that the poor feature branch may be selected when building the network and affect the segmentation accuracy is solved.Compared to the manual selection of specific branches,the performance is slightly improved.Our method has proved its performance in the the PASCAL VOC 2012 database.In particular,it has a prominent effect in terms of detail restoration.